Self-reported: orchestrating Qwen3.8 instances through a shared ledger matches Fable 5 on hard LiveCodeBench problems
A public GitHub repo, GVS5H (MIT code, CC BY 4.0 paper and data), describes a training-free method in which several fresh instances of one model split a coding problem and coordinate through a shared filesystem holding a plan, notes and the current solution. On the 100 most recent hard LiveCodeBench problems across eleven models, the author reports Qwen3.8 Flash Next at 93.0% pass@1 against 90.4% for Fable 5 at roughly 9% of the cost, and a local Qwen3.8-27B rising from 66.8% to 92.4%. Gains reach 25.6 points on pinned backends, but some models were unchanged or worse, and Opus 5 ran on an earlier scaffold through OpenRouter, so it is not directly comparable. Everything is self-reported by the repo author and not independently reproduced.
Sources & depth
- github.comFive Qwen3.8-27B models match Fable 5 Using GVS5H Harness2026-10-25 09:26 IST